Information Retrieval- Query expansion

Jian-Yun Nie

1

Recap of previous lectures

Indexing Traditional IR models Lemur toolkit

How to run it How to modify it (for query expansion)

Evaluating a search engine Benchmarks Measures

2

This lecture

Improving results For high recall. E.g., searching for aircraft didn’t

match with plane; nor thermodynamic with heat Options for improving results…

Focus on relevance feedback The complete landscape

Global methods Query expansion

Thesauri Automatic thesaurus generation

Local methods Relevance feedback Pseudo relevance feedback 3

Review of traditional IR models

Document and query are represented by a set of terms, organized in some way (vector, probabilistic model, …)

Preprocessing on words: stemming to create the same term for related words

Each indexing term is considered to represent a unique meaning

4

Assumptions

1. Different terms are assumed to represent different meanings

phone vs. telephone Information retrieval vs. search engine Consequence: silence – relevant documents are

not retrieved

2. A term is assumed to represent only one meaning

table: furniture, data structure, … office: a work place, an organization, software, Consequence: noise: irrelevant documents are

retrieved 5

Possible ways to deal with the first problem

Different terms may represent the same meaning Create a semantic representation

Each term is mapped to a concept Two terms representing the same meaning are mapped to the

same concept Problem: requires extensive semantic resources – not feasible at

large scale now (may be done in specialized area such as medicine)

Using relationships between terms in retrieval Term b means the same thing as term a （ b a) Query a ：

Match documents containing a Match documents containing b

Equivalent to consider a query b∨a (Query expansion) If b is only related to a (not the same meaning), one may want to

decrease the weight of b in the query6

Possible ways to deal with the second problem

A term may mean different things (ambiguity) Semantic representation

A term is mapped to different concepts depending on what it means

Term disambiguation Often difficult to do, and the experiments using word

sense disambiguation has not proven to be effective Use compound terms/phrases instead of single

terms Office update Office address Q: Does this help in practice?

We will come back on this later 7

Query expansion Goal: extend the initial query by adding related terms E.g. phone number phone number, telephone Why is it necessary to expand query?

Queries are short: 2-3 words They do not include all the words that may describe the information

need They only describe some of the aspects of the information need

Can we automatically complete the query so as to arrive at a better and more complete description of the information need?

Key problems Recognize term relationships: phonetelephone Determine how strongly the new term is related to the query Combine the new terms with the initial query 8

Term relationships

Various relationships between terms: Syntactic:

ADJ-NN (e.g. beautiful campus) Lexical

NN ADJ (e.g. computation computational) Semantic:

Synonymy: computer ↔ electronic computer Hypernymy: computer machine Hyponymy: machine computer Related to: program computer …

What relationships are useful for IR? 9

Term relationships

Useful relationships for IR When one asks for documents on a, a document

on b can also be relevant ab is useful

Usually, relationships are defined between terms Assumption: a query expanded by related terms is

a related query, and the documents matching the expanded query also match the initial query

10

How to determine term relationships?

Thesaurus A thesaurus contains a set of manually defined

relations between terms Synonymy, hypernymy, hyponymy, meronymy,

holonymy, … Term co-occurrences in documents

Two terms that co-occur often are related Relevance feedback

Terms extracted from the relevant documents are related to the query

11

Example ： Wordnet

relationship example

synonymy computer data processor

antonymy big small

hyponymy

(A-KIND-OF)

tree hyponymy maple

hypernymy (IS-A) maple hypernymy tree

meronymy (HAS) computer meronymy processor

holonymy

(IS-PART-OF)

processor holonymy computer

Relevance Feedback

Relevance feedback: user feedback on relevance of docs in initial set of results User issues a (short, simple) query The user marks returned documents as relevant or

non-relevant. The system computes a better representation of the

information need based on feedback. Relevance feedback can go through one or more

iterations. Idea: it may be difficult to formulate a good query

when you don’t know the collection well, so iterate13

Relevance Feedback: Example

Image search engine http://nayana.ece.ucsb.edu/imsearch/imsearch.html

14

Results for Initial Query

15

Relevance Feedback

16

Results after Relevance Feedback

17

Rocchio Algorithm The Rocchio algorithm incorporates relevance

feedback information into the vector space model. Want to maximize sim (Q, Cr) - sim (Q, Cnr) The optimal query vector for separating relevant

and non-relevant documents (with cosine sim.):

Qopt = optimal query; Cr = set of rel. doc vectors; N = collection size

Unrealistic: we don’t know relevant documents.

rjrj Cdj

rCdj

ropt d

CNd

CQ

11

18

The Theoretically Best Query

x

x

xx

oo

o

Optimal query

x non-relevant documentso relevant documents

o

o

o

x x

xxx

x

x

x

x

x

x

x

x

x

19

Rocchio 1971 Algorithm (SMART)

Used in practice:

qm = modified query vector; q0 = original query vector; α,β,γ: weights (hand-chosen or set empirically); Dr = set of known relevant doc vectors; Dnr = set of known irrelevant doc vectors

New query moves toward relevant documents and away from irrelevant documents

Tradeoff α vs. β/γ : If we have a lot of judged documents, we want a higher β/γ.

Term weight can go negative Negative term weights are ignored

nrjrj Ddj

nrDdj

rm d

Dd

Dqq

110

20

Relevance feedback on initial query

x

x

xx

oo

o

Revised query

x known non-relevant documentso known relevant documents

o

o

ox

x

x x

xx

x

x

xx

x

x

x

x

Initial query

21

Relevance Feedback in vector spaces

We can modify the query based on relevance feedback and apply standard vector space model.

Use only the docs that were marked. Relevance feedback can improve recall and

precision Relevance feedback is believed to be most useful

for increasing recall in situations where recall is important Users can be expected to review results and to take

time to iterate22

Positive vs Negative Feedback

Positive feedback is more valuable than negative feedback (so, set < ; e.g. = 0.25, = 0.75).

Many systems only allow positive feedback (=0).

Why?

23

Probabilistic relevance feedback

Rather than reweighting in a vector space… If user has told us some relevant and irrelevant

documents, then we can proceed to build a classifier, such as a Naive Bayes model: P(tk|R) = |Drk| / |Dr| P(tk|NR) = (Nk - |Drk|) / (N - |Dr|)

tk = term in document; Drk = known relevant doc containing tk; Nk = total number of docs containing tk

Cf. classification This is effectively another way of changing the

query term weights But note: the above proposal preserves no

memory of the original weights24

Relevance Feedback: Assumptions

A1: User has sufficient knowledge for initial query. A2: Relevance prototypes are “well-behaved”.

Term distribution in relevant documents will be similar

Term distribution in non-relevant documents will be different from those in relevant documents

Either: All relevant documents are tightly clustered around a single prototype.

Or: There are different prototypes, but they have significant vocabulary overlap.

Similarities between relevant and irrelevant documents are small 25

Violation of A1

User does not have sufficient initial knowledge. Examples:

Misspellings (Brittany Speers). Cross-language information retrieval (hígado). Mismatch of searcher’s vocabulary vs. collection

vocabulary Cosmonaut/astronaut

26

Violation of A2

There are several relevance prototypes. Examples:

Burma/Myanmar Contradictory government policies Pop stars that worked at Burger King

Often: instances of a general concept Good editorial content can address problem

Report on contradictory government policies

27

Relevance Feedback: Problems

Long queries are inefficient for typical IR engine. Long response times for user. High cost for retrieval system. Partial solution:

Only reweight certain prominent terms Perhaps top 20 by term frequency

Users are often reluctant to provide explicit feedback

It’s often harder to understand why a particular document was retrieved after apply relevance feedback

Why?

28

Relevance Feedback Example: Initial Query and Top 8 Results

Query: New space satellite applications

+ 1. 0.539, 08/13/91, NASA Hasn't Scrapped Imaging Spectrometer

+ 2. 0.533, 07/09/91, NASA Scratches Environment Gear From Satellite Plan

3. 0.528, 04/04/90, Science Panel Backs NASA Satellite Plan, But Urges Launches of Smaller Probes

4. 0.526, 09/09/91, A NASA Satellite Project Accomplishes Incredible Feat: Staying Within Budget

5. 0.525, 07/24/90, Scientist Who Exposed Global Warming Proposes Satellites for Climate Research

6. 0.524, 08/22/90, Report Provides Support for the Critics Of Using Big Satellites to Study Climate

7. 0.516, 04/13/87, Arianespace Receives Satellite Launch Pact From Telesat Canada

+ 8. 0.509, 12/02/87, Telecommunications Tale of Two Companies

Note: want high recall

29

Relevance Feedback Example: Expanded Query

2.074 new 15.106 space 30.816 satellite 5.660 application 5.991 nasa 5.196 eos 4.196 launch 3.972 aster 3.516 instrument 3.446 arianespace 3.004 bundespost 2.806 ss 2.790 rocket 2.053 scientist 2.003 broadcast 1.172 earth 0.836 oil 0.646 measure

30

Top 8 Results After Relevance Feedback

+ 1. 0.513, 07/09/91, NASA Scratches Environment Gear From Satellite Plan

+ 2. 0.500, 08/13/91, NASA Hasn't Scrapped Imaging Spectrometer

3. 0.493, 08/07/89, When the Pentagon Launches a Secret Satellite, Space Sleuths Do Some Spy Work of Their Own

4. 0.493, 07/31/89, NASA Uses 'Warm‘ Superconductors For Fast Circuit

+ 5. 0.492, 12/02/87, Telecommunications Tale of Two Companies

6. 0.491, 07/09/91, Soviets May Adapt Parts of SS-20 Missile For Commercial Use

7. 0.490, 07/12/88, Gaping Gap: Pentagon Lags in Race To Match the Soviets In Rocket Launchers

8. 0.490, 06/14/90, Rescue of Satellite By Space Agency To Cost $90 Million

31

Evaluation of relevance feedback strategies

Use q0 and compute precision and recall graph

Use qm and compute precision recall graph Assess on all documents in the collection

Spectacular improvements, but … it’s cheating! Partly due to known relevant documents ranked higher Must evaluate with respect to documents not seen by user

Use documents in residual collection (set of documents minus those assessed relevant)

Measures usually then lower than for original query But a more realistic evaluation Relative performance can be validly compared

Empirically, one round of relevance feedback is often very useful. Two rounds is sometimes marginally useful.

32

Relevance Feedback on the Web

Some search engines offer a similar/related pages feature (this is a trivial form of relevance feedback)

Google (link-based) Altavista Stanford WebBase

But some don’t because it’s hard to explain to average user: Alltheweb msn Yahoo

Excite initially had true relevance feedback, but abandoned it due to lack of use.

α/β/γ ??

33

Excite Relevance Feedback

Spink et al. 2000 Only about 4% of query sessions from a user

used relevance feedback option Expressed as “More like this” link next to each

result But about 70% of users only looked at first page

of results and didn’t pursue things further So 4% is about 1/8 of people extending search

Relevance feedback improved results about 2/3 of the time

34

Other Uses of Relevance Feedback

Following a changing information need Maintaining an information filter (e.g., for a news

feed) User clicks as relevance feedback [Shen et al. 05] Active learning

[Deciding which examples it is most useful to know the class of to reduce annotation costs]

35

Relevance FeedbackSummary

Relevance feedback has been shown to be very effective at improving relevance of results. Requires enough judged documents, otherwise it’s

unstable (≥ 5 recommended) Requires queries for which the set of relevant

documents is medium to large Full relevance feedback is painful for the user. Full relevance feedback is not very efficient in most

IR systems. Other types of interactive retrieval may improve

relevance by as much with less work.36

Pseudo-Relevance Feedback

As true relevance feedback is hard to obtain, we assume that the top-ranked documents in the initial retrieval results are relevant

Then we use the same query modification process to create a new query

Notice that this subset of documents are not all relevant. However, they are often more relevant than the documents at lower ranks. So, they still capture some characteristics of relevance.

37

Pseudo-Relevance Feedback in VSM Vector space model

DF= set of top-ranked documents Usually, we also select the k strongest terms from top-

ranked documents, i.e. only keep the k terms in and let the other terms to be 0.

A typical figure: use top 10-20 documents, and the 20-100 strongest terms.

There are experiments with massive expansion, e.g. using 500 terms from 100 top documents (Massive QE – Buckley et al.) . But this is unrealistic. 38

Pseudo-Relevance Feedback in LM

KL divergence:

39

40

Expanding query model

Classical LM

Feedback model

Estimating the feedback model

Relevance Model (Lavrenko and Croft, 2001) Viewing top-ranked documents as relevance samples

Mixture Model (Zhai and Lafferty, 2001) Generating top-ranked documents from two sources: topic

model and general model (collection) Fitting topic model with EM algorithm by maximizing the

likelihood of top-ranked documents

41

Feedback as Model Interpolation

Query Q

D

)||( DQD

Document D

Results

Feedback Docs F={d1, d2 , …, dn}

FQQ )1('

Generative model

Q

F=0

No feedback

FQ '

=1

Full feedback

QQ '

42

Mixture model – estimating feedback model by EM

Principle of Expectation Maximization (EM): Given a set of documents, determine a component model so that the global model can maximize the likelihood of the set of documents

Global model (documents’ likelihood) to maximize

E-step

M-step43

Results with mixture feedback model

Divergence minimization is another model based on feedback documents (see Zhai and Lafferty 2001)

44

Results with Relevance model

45

The complete landscape

Local methods Relevance feedback Pseudo relevance feedback

Variant: use passage retrieval, and top-retrieved passages for feedback (better than document feedback)

Passage = section or overlapping text window (e.g. 200 words)

(Why?)

Global methods Query expansion/reformulation

Thesauri (e.g. WordNet, HowNet) Automatic thesaurus generation 46

Query Expansion

In relevance feedback, users give additional input (relevant/non-relevant) on documents, which is used to extract additional terms for query

In interactive query expansion, users give additional input (good/bad search term) on words or phrases.

47

Query Expansion: Example

Also: see www.altavista.com, www.teoma.com

48

Controlled Vocabulary

49

Types of Query Expansion Global Analysis: Thesaurus-based

Controlled vocabulary Maintained by editors (e.g., medline)

Manual thesaurus E.g. MedLine: physician, syn: doc, doctor, MD, medico,

WordNet Automatically derived thesaurus

(co-occurrence statistics) Refinements based on query log mining

Common on the web

Local Analysis (better than Global analysis –Xu and Croft): Analysis of documents in result set 50

QE based on thesauri

Thesaurus Relations between terms E.g. WordNet (http://wordnetweb.princeton.edu/perl/webwn)

51

Thesaurus-based Query Expansion

This doesn’t require user input For each term, t, in a query, expand the query with

synonyms and related words of t from the thesaurus feline → feline cat

May weight added terms less than original query terms. Generally increases recall. Widely used in many science/engineering fields May significantly decrease precision, particularly with

ambiguous terms. “interest rate” “interest rate fascinate evaluate”

There is a high cost of manually producing a thesaurus And for updating it for scientific changes

52

Problems using Wordnet Strength: the stored relations are manually validated However:

Coverage: not all the terms are included Usefulness: not all the related terms are useful for IR (e.g.

computercomputing machine} Ambiguity: computer machine? Or expert? Lack of weighting: strong and weak relations are not

distinguished Experiments

(Voorhees 1993, 1994): Wordnet does not help Automatically adding related terms decreases effectiveness Even adding correct related synsets does not help

Others: some improvements using appropriate weighting (according to collection statistics) 53

Experiments of Voorhees

1993 For a query, determine the most relevant synsets Create an additional vector using synset ids Or additional vectors using related terms/synsets Combining similarities Does not improve over bag-of-words vector

1994 Is it due to ambiguity? Manual selection of synsets Does not improve

54

Automatic extraction of term relationships

Co-occurrence: two terms occur at the same time within the same text (fragment)

Assumption: The more two terms co-occur, the stronger is the relationship between them.

Two aspects to consider: The context in which co-occurrences are considered:

document, paragraph, sentence, passage, text window (e.g. 10 words)…

Calculation of the strength: f(ab)=#co-occ(a,b) / #occ(a) or #co-occ(a,b) / Σc#co-occ(a,c) f(ab)=|A∩B| / |A B|∪ A=set of contexts of a Pointwise mutual information MI(a,b)=P(a,b)/P(a)*P(b)

55

Automatic Thesaurus GenerationExample

56

Analysis of co-occurrence relations Coverage: large Nature: no semantic nature can be determined, but

usually suggest that the two terms are used in the same contexts, to describe the same topics

Problems Noise: semantically unrelated terms can co-occur Silence: related terms may not co-occur often (tyre-tire) Co-occ. are domain dependent: usually cannot use the

results obtained from one area to another area. Strength

Weighting Co-occurring terms can correspond to some related topics

(may be relevant) 57

Comparison between co-occ and WordNet

Cao et al. 05, Mandala et al. 98/99 Co-occurrence relations work better than WordNet

relations Using co-occurrences to etimate the strength of

relations in WordNet Show improvements It is important to weight the relations

58

How to integrate term relationships?

Vector space modelVoorhees (1993, 1994)

The expansion terms form another query vector Similarities with the original vector and the expansion vector

are interpolatedQiu and Frei (1993)

Determine a relation of the term to the whole query Sum up its relations to all the query terms

qi: weight of ti in q Expect to reduce noise and ambiguity:

terms related to one query term may not be strongly related to the query

Q: Is this true?59

Analysis of Qiu and Frei If sim(report, research) and sim(report, paper)

high

vs sim(sheet, research) low and sim(sheet, paper) high;

Then sim(report, research paper) > sim(sheet, research paper)

Howeversim(programming, Java) can be very highSo sim(programming, Java island) can be quite highShould we conclude that programming is related to Java island?

60

Analysis of Qiu and Frei What’s wrong? Relations between terms are extracted by

considering terms in isolation Problem of ambiguity

Summing up the similarities with all the query terms can promote some truly related terms

But this does not solve the fundamental problem that the initial term-term relations are context-independent

We have no information on when sim(programming, Java) applies

Solution: Context-dependent relations (see Bai et al. 06) – determine sim(programming, Java+language)

61

How to integrate term relationships?

62

An example (from Bai et al. 06)

63

UM- classical unigram LMUQE- unigram query expansionBQE- bigram query expansion

Query Reformulation/suggestion

Suggestions Thesaurus Controlled vocabulary Browse lists of terms in the inverted index Effective resources for search engines

Query language model (what words are the next?) Query frequency User query sessions (a query is reformulated into another) Clickthrough: A query which generates many clicks is a

successful one (a good candidate to use in reformulation)

Users use suggestions, but not feedback. Why?64

Query Expansion: Summary

Query expansion is often effective in increasing recall. Not always with general thesauri Fairly successful for subject-specific collections

Usually, query expansion is considered as a means to increase recall, and it hurts precision. This is not entirely true. Why?

Overall, not as useful as relevance feedback; may be as good as pseudo-relevance feedback

65

66

Query expansion v.s. document expansion

Document expansion

Query expansion

Query expansion vs. document expansion (Cao et al. 2007)

It is observed that when the same resource (e.g. term co-occ. statistics) are used for query expansion and document expansion, query expansion is more effective. Why?

But document expansion can be performed offline !

67

Document expansion in search engines

Document descriptions Title (very useful) Body (useful) Anchor texts pointing to the document (useful) Queries leading to click on the document (useful)

68

Title: IRDody: …

Search engine technology

Document retrieval

AnchorQuery•IR•Indexing•IR literature•…

Questions for discussion (next lecture) Why is it necessary to perform query expansion? Why isn’t is so successful in

practice even though it may increase retrieval effectiveness? Compare and contrast different ways to do query expansion What is the difference between global query expansion (or global context

analysis) and local query expansion (local context analysis)? Why did some experiments show better performances with local query expansion? (see Xu and Croft 96)

Are the strong terms extracted from feedback documents all useful? If not, how to select? (see Cao et al. 06)

Term relationships extracted are used in a context-independent way (i.e. one term is considered related to another whatever the query). What are the possible problems this may bring? What are your solutions? (see Bai et al. 07)

Do you have a way to make query expansion useable by end users? We talked about query expansion using feedback documents, thesauri and co-

occurrence statistics. What other methods/resources can you think of to do query expansion? (see Carpineto’s survey)

Does query expansion allows us to move IR from term matching to sense matching?

Read Carpineto’s survey on query expansion. 69

ResourcesMG Ch. 4.7MIR Ch. 5.2 – 5.4Yonggang Qiu , Hans-Peter Frei, Concept based query expansion. SIGIR

16: 161–169, 1993.Schuetze: Automatic Word Sense Discrimination, Computational Linguistics,

1998.Singhal, Mitra, Buckley: Learning routing queries in a query zone, ACM

SIGIR, 1997.Buckley, Singhal, Mitra, Salton, New retrieval approaches using SMART:

TREC4, NIST, 1996.Gerard Salton and Chris Buckley. Improving retrieval performance by

relevance feedback. Journal of the American Society for Information Science, 41(4):288-297, 1990.

Bai, J. Nie, J.-Y., Bouchard,H., Cao, G. Using Query Contexts in Information Retrieval, SIGIR, pp. 15-22, 2007.

Cao G. Nie, J.Y., Bai, J., Using Markov Chains to Exploit Word Relationships in Information Retrieval, RIAO, 2007

G. Cao, J.Y. Nie, J. Gao, S. Robertson, Selecting Good Expansion Terms for Pseudo-Relevance Feedback, ACM-SIGIR, 2008, pp. 243-250

70

ResourcesHarman, D. (1992): Relevance feedback revisited. SIGIR 15: 1-10Chris Buckley, Gerard Salton, and James Allan.

The effect of adding relevance information in a relevance feedback environment. In SIGIR 17, pages 292-300, Dublin, Ireland, 1994.

Xu, J., Croft, W.B. (1996): Query Expansion Using Local and Global Document Analysis, in SIGIR 19: 4-11.

Spink, A., Jansen, J. and Ozmultu, H.C. (2000) "Use of query reformulation and relevance feedback by Excite users." Internet Research: Electronic Networking Applications and Policy. http://ist.psu.edu/faculty_pages/jjansen/academic/pubs/internetresearch2000.pdf

Zhai, C., Lafferty, J., Model-based feedback in the language modeling approach to information retrieval, CIKM 2001.

Berger, A., Lafferty, J., Information retrieval as statistical translation, SIGIR 1999, pp. 222-229.

Xuehua Shen, Bin Tan, ChengXiang Zhai: Implicit user modeling for personalized search. CIKM 2005: 824-831 71